Swap errors and steps number, refactor in merge_steps function, use n…

…ew algorithm for parallel validation.

AIAgent/common/constants.py

@@ -47,4 +47,6 @@ class ResultsHandlerLinks:
 BASE_NN_OUT_FEATURES_NUM = 8
 
 # assuming we start from /VSharp/VSharp.ML.AIAgent
-SERVER_WORKING_DIR = "../GameServers/VSharp/VSharp.ML.GameServer.Runner/bin/Release/net7.0/"
+SERVER_WORKING_DIR = (
+ "../GameServers/VSharp/VSharp.ML.GameServer.Runner/bin/Release/net7.0/"
+)

AIAgent/learning/play_game.py

@@ -99,8 +99,8 @@ def add_single_step(input, output):
  map_result = (
  model_result.actual_coverage_percent,
  -model_result.tests_count,
- model_result.errors_count,
  -model_result.steps_count,
+ model_result.errors_count,
  )
  with_dataset.update(with_connector.map.MapName, map_result, map_steps)
  return model_result, end_time - start_time

AIAgent/ml/common_model/dataset.py

@@ -3,6 +3,7 @@
 
 import os
 import numpy as np
+import random
 
 import tqdm
 import logging
@@ -91,7 +92,11 @@ def get_plain_data(self, threshold: int = 100):
  result = []
  for map_result, map_steps in self.maps_data.values():
  if map_result[0] >= threshold:
- for step in map_steps:
+ if len(map_steps) > 2000:
+ selected_steps = random.sample(map_steps, 2000)
+ else:
+ selected_steps = map_steps
+ for step in selected_steps:
  if step.use_for_train:
  result.append(step)
  return result
@@ -125,12 +130,15 @@ def update(
  x.to("cpu")
  filtered_map_steps = self.filter_map_steps(map_steps)
  if map_name in self.maps_data.keys():
- if self.maps_data[map_name][0] == map_result:
- logging.info(
- f"Steps on map {map_name} were merged with current steps with result {map_result}"
- )
+ if self.maps_data[map_name][0] == map_result and map_result[0] == 100:
+ init_steps_num = len(self.maps_data[map_name][1])
+
  filtered_map_steps = self.remove_similar_steps(filtered_map_steps)
  self.merge_steps(filtered_map_steps, map_name)
+ new_steps_num = len(self.maps_data[map_name][1])
+ logging.info(
+ f"Steps on map {map_name} were merged with current steps with result {map_result}. {len(filtered_map_steps)} + {init_steps_num} -> {new_steps_num}. "
+ )
  if self.maps_data[map_name][0] < map_result:
  logging.info(
  f"The model with result = {self.maps_data[map_name][0]} was replaced with the model with "
@@ -177,21 +185,28 @@ def flatten_and_sort_hetero_data(step: HeteroData):
  old_steps = create_dict(self.maps_data[map_name][1])
 
  for vertex_num in new_steps.keys():
+ flattened_old_steps = []
+ if vertex_num in old_steps.keys():
+ for old_step in old_steps[vertex_num]:
+ flattened_old_steps.append(flatten_and_sort_hetero_data(old_step))
  for new_step in new_steps[vertex_num]:
- if vertex_num in old_steps.keys():
- for step_num, old_step in enumerate(old_steps[vertex_num]):
- new_g_v, new_s_v = flatten_and_sort_hetero_data(new_step)
- old_g_v, old_s_v = flatten_and_sort_hetero_data(old_step)
-
- if np.array_equal(new_g_v, old_g_v) and np.array_equal(
- new_s_v, old_s_v
- ):
- y_true_sum = old_step.y_true + new_step.y_true
- y_true_sum[y_true_sum != 0] = 1
-
- old_step.y_true = y_true_sum / torch.sum(y_true_sum)
- old_steps[vertex_num][step_num] = old_step
- break
- merged_steps.append(new_step)
+ new_g_v, new_s_v = flatten_and_sort_hetero_data(new_step)
+ should_add = True
+ for step_num, (old_g_v, old_s_v) in enumerate(flattened_old_steps):
+ if np.array_equal(new_g_v, old_g_v) and np.array_equal(
+ new_s_v, old_s_v
+ ):
+ y_true_sum = (
+ old_steps[vertex_num][step_num].y_true + new_step.y_true
+ )
+ y_true_sum[y_true_sum != 0] = 1
+
+ old_steps[vertex_num][step_num].y_true = y_true_sum / torch.sum(
+ y_true_sum
+ )
+ should_add = False
+ break
+ if should_add:
+ merged_steps.append(new_step)
  merged_steps.extend(sum(old_steps.values(), []))
  self.maps_data[map_name] = (self.maps_data[map_name][0], merged_steps)

AIAgent/run_common_model_training.py

@@ -145,19 +145,16 @@ def train(trial: optuna.trial.Trial, dataset: FullDataset):
 
  with game_server_socket_manager() as ws:
  all_maps = get_maps(websocket=ws, type=MapsType.TRAIN)
- maps = np.array_split(all_maps, GeneralConfig.SERVER_COUNT)
- random.shuffle(maps)
  tasks = [
- (maps[i], FullDataset("", ""), cmwrapper)
- for i in range(GeneralConfig.SERVER_COUNT)
+ ([all_maps[i]], FullDataset("", ""), cmwrapper)
+ for i in range(len(all_maps))
  ]
 
  mp.set_start_method("spawn", force=True)
- # p = Pool(GeneralConfig.SERVER_COUNT)
 
  all_average_results = []
  for epoch in range(config.epochs):
- data_list = dataset.get_plain_data()
+ data_list = dataset.get_plain_data(80)
  data_loader = DataLoader(data_list, batch_size=config.batch_size, shuffle=True)
  print("DataLoader size", len(data_loader))
 
@@ -194,7 +191,7 @@ def train(trial: optuna.trial.Trial, dataset: FullDataset):
  cmwrapper.make_copy(str(epoch + 1))
 
  with mp.Pool(GeneralConfig.SERVER_COUNT) as p:
- result = list(p.map(play_game_task, tasks))
+ result = list(p.map(play_game_task, tasks, 1))
 
  all_results = []
  for maps_result, maps_data in result:
@@ -228,7 +225,6 @@ def train(trial: optuna.trial.Trial, dataset: FullDataset):
  torch.save(model.state_dict(), Path(path_to_model))
  del data_list
  del data_loader
- # p.close()
 
  return max(all_average_results)